Upload README.md with huggingface_hub

README.md

@@ -59,10 +59,8 @@ boxes**, and **a caption for every box**. It was built by [Breakpoint AI](https:
 to train a 10B-parameter joint image + bounding-box diffusion model for synthetic object-detection
 data generation.
 
-<!-- Once the example figure is ready, uncomment and upload it (see update_hub_card.py --asset):
 ![An image with bounding boxes drawn over detected objects, each box captioned, alongside the
 image's global caption](./example_figure.png)
--->
 
 - **54,909,477 images**, each with one image-level caption
 - **511,652,664 bounding boxes** (9.32 per image on average), each with its own caption
@@ -79,9 +77,9 @@ generated images and bounding boxes simultaneously. Given 10 labeled images, it
 new scene and start producing labeled training images.
 
 In academia, there has been considerable research on conditional generation (bounding boxes to
-image) <!-- TODO: cite GLIGEN, InstanceDiffusion, etc. -->, and the associated datasets
-<!-- TODO: cite -->. While helpful, this work did not touch joint generation, and unfortunately
-was not at the scale necessary to be actually useful.
+image) (e.g. GLIGEN [Li et al., 2023], InstanceDiffusion [Wang et al., 2024]), and the associated
+datasets (e.g. ROVI [Peng et al., 2025]). While helpful, this work did not touch joint generation,
+and unfortunately was not at the scale necessary to be actually useful.
 
 In order to get a better performing model, we decided to build this dataset, which is over 50
 times larger than comparable datasets (e.g. ROVI, ICCV 2025, at 1M images). Having this much data
@@ -184,23 +182,34 @@ the image) and the affected rows will be removed.
 
 ## Appendix: Annotation pipeline
 
-<!-- TODO: fill in — which object detector (Jamie mentioned Grounding DINO), which captioning
-     model(s), and the prompts used for each. Franklin has this detail. -->
+- **Object detection**: bounding boxes and class labels were produced by an open-vocabulary
+  detector (Grounding DINO) as an upstream step.
+- **Per-object captions**: [InternVL3-2B](https://huggingface.co/OpenGVLab/InternVL3-2B), served
+  via LMDeploy (TurboMind engine, `internvl2_5` chat template). Each detected box was cropped
+  (no padding) and captioned with the prompt:
+  > `<image>` Describe this `<category name>` in 20 words or less and focus on its color,
+  > appearance, size, and style, etc.
+
+  where `<category name>` is the detector's class label for that box. Generated with
+  `max_new_tokens=64`.
+- **Image-level captions**: [InternVL3-8B](https://huggingface.co/OpenGVLab/InternVL3-8B) (via
+  `transformers`), prompted with:
+  > `<image>` Write a single, natural-language description of this image. Do not use bullet
+  > points or lists. Keep it under 100 words.
+
+  Generated with `max_new_tokens=128`, sampling enabled.
 
 ## Citation
 
 ```bibtex
 @misc{breakpoint_grounding_55m,
   title  = {Breakpoint Grounding 55M},
-  author = {Wang, Franklin and {[Desmond LASTNAME]} and Murdoch, Jamie},
+  author = {Wang, Franklin and Zhong, Desmond and Murdoch, Jamie},
   year   = {2026},
   url    = {https://huggingface.co/datasets/BreakpointAI/breakpoint-grounding-55m}
 }
 ```
 
-<!-- TODO: confirm Desmond's last name and Jamie's preferred citation name (Jamie Murdoch vs.
-     William Murdoch), and author order (currently Franklin, Desmond, Jamie per Slack). -->
-
 ### Related work
 
 ```bibtex